Asaad Mohammedsaleh earns MSc on genome-scale protein function adjustment
Asaad Mohammedsaleh has completed his MSc in Computer Science with a thesis on a constraint-optimization framework that repairs the consistency of genome-scale protein function predictions.
About
Asaad Mohammedsaleh has successfully defended his MSc thesis, Genome-Scale Protein Function Adjustment using Constraint Optimization, at KAUST, under the supervision of Professor Robert Hoehndorf. The work tackles a known weakness of state-of-the-art protein function prediction methods: they perform well on individual proteins but degrade noticeably when evaluated across whole proteomes, especially on consistency metrics such as taxon coherence and complex-completion.
Asaad's contribution is a predictor-agnostic post-processing step built on the OR-Tools CP-SAT solver. It applies two sequential stages: a taxon-consistency stage that removes Gene Ontology term predictions which violate an organism's taxonomic constraints, followed by a complex-coherence stage that repairs singleton complex annotations. The optimization is framed as an adjustment problem: starting from the predictions produced by any single-protein method, the solver makes minimal modifications to satisfy genome-scale biological constraints while preserving the original predictive signal.
In evaluations on 568 UniProt reference proteomes across two held-out splits, the adjustment framework was applied on top of four baselines (Seq-Sim, MLP-ESM2, DeepGO-SE, and SPROF-GO). The per-protein scores (F-max, S-min, AUC, AUPR) were preserved across all methods, while taxon consistency and complex coherence reached 100%. The solver runs over proteomes of tens of thousands of proteins in seconds, making it practical to bolt onto existing prediction pipelines.
The result demonstrates that a flexible, predictor-agnostic post-processor can produce system-level, biologically consistent annotations on top of any existing method, without sacrificing per-protein accuracy. Congratulations, Asaad!